Construction and validation of image-based statistical shape and intensity models of bone

Three-dimensional models of bone structures, inferred from 3D and 2D imaging modalities, provide a number of uses for medical professionals. Such models are typically constructed using mesh-based approaches and some form of principal component analysis (PCA). Image-based approaches are understood to have a number of advantages over mesh-based ones, such as the absence of intrinsic segmentation as well as a better reproduction fidelity. Furthermore, Gaussian process morphable models have recently been shown to offer added benefits to the traditional PCA-based approach, such as the ability to easily incorporate prior knowledge into the model, and a resolution that can be made application-specific. We demonstrate how to build an image-based statistical shape and intensity model (SSIM) using Gaussian process morphable models and validate the quality of the model using common mesh model metrics that we adapt to the image-based paradigm. Our results show that the model can generate valid novel femur examples and can be registered to unseen femur examples with average root mean square error and average mutual information score of 0.172 and 0.644, respectively. However, only twenty binary training examples are used, each limited to contain approximately 65000 voxels. Future work will aim at extending the image-based SSIM to include the full range of CT intensity values; larger CT volumes and improve on the model building time.